Skip to main content
This is the scenario the entire course has been building toward. In scenario 4, we generated a styled invoice PDF for “Northwind Studio” by writing a long, careful prompt. It worked. But anyone who actually sends invoices is going to send another one next week, and another the week after. The prompt becomes copy-paste, then it becomes a Word doc someone shares around, and eventually someone says “this should just be a function.” That last step is what /skill-creator does. It takes a recurring prompt and turns it into a named, parameterized skill the model can call by name forever after — same input shape, same output shape, no prompt engineering required.

​
The prompt

I keep generating styled invoice PDFs from the same template (header band,
monospace numbers, line items, subtotal/tax/total).

Use the skill-creator skill to package this into a reusable skill called
"invoice-builder" that takes structured input (client, invoice number,
line items, tax rate) and produces a styled PDF.

Then demonstrate the new skill by calling it once with sample data:
client "Globex Corp", invoice 2026-0099, two line items
(Consulting 10h x $150, Implementation $2500), tax 7%.
The skill-creator prompt entered into chat.yambr.com

​
What happened

The skill-creator skill is itself a skill. Its job is to walk the model through producing a well-formed skill folder: SKILL.md describing capabilities, scripts implementing the work, an __init__.py so the skill is callable. On this run, you can see the model building a tool, then using the tool it just built: The model methodically reading skill-creator and pdf, then building the invoice-builder skill step by step Step by step, in order:
  1. Explored view — read the skill-creator guide to understand the format.
  2. Explored view — read the pdf skill for reference, since that’s what invoice-builder will use under the hood.
  3. Explored bash_tool — initialized the new skill.
  4. Explored bash_tool — created the skill in the workspace.
  5. Explored create_file — wrote the Python script that generates styled invoices.
  6. Explored str_replace — fixed a syntax error live.
  7. Explored bash_tool — tested the script.
  8. Explored bash_tool — viewed the result to verify the styling.
  9. Explored create_file — created SKILL.md.
  10. View Result from create_file — created __init__.py to make it a proper Python module.
Then the model called the brand-new skill with the demo data from the prompt.

​
The output

Demonstration invoice generated by the new invoice-builder skill — usage example code on the left, INVOICE 2026-0099 PDF on the right Two things are visible in the final reply: 1. The usage contract. Right in the chat, as code: 
build_invoice(
    output_path="invoice.pdf",
    client="Globex Corp",
    invoice_number="2026-0099",
    line_items=[
        {"description": "Consulting", "quantity": 10, "rate": 150, "unit": "hours"},
        {"description": "Implementation", "amount": 2500},
    ],
    tax_rate=0.07,  # 7%
)
That’s the skill’s surface. Anyone — the same user tomorrow, a teammate, an n8n workflow, a CI job — can call it with that exact shape. 2. The demonstration invoice in the right-hand panel: a real PDF with a blue “INVOICE 2026-0099” header band, “Client: Globex Corp,” the two line items priced correctly (1,500.00and1,500.00 and 2,500.00), Tax (7%) of 280.00,and∗∗Total:280.00, and **Total: 4,280.00**. The math is computed inside the skill — not by the LLM.

​
What just happened (the leverage)

Compare the two prompts side by side: Before (scenario 4):
Generate a styled invoice as a PDF for a freelance design studio called “Northwind Studio”. Bill to: Acme Corp, 123 Main St, San Francisco, CA. Invoice #2026-0042, dated 2026-04-26, due 2026-05-26. Line items: Logo redesign (2,400),Brandguidelinesdoc(2,400), Brand guidelines doc (1,200), 3 x landing page mockup (800 each), Asset handoff & support (600). Subtotal, 8.5% tax, and total. Use a clean modern layout with the studio name in a colored header band, monospace numbers, and a thin accent line. Use the pdf skill.
That’s 73 words and a careful structure that has to be re-typed (or copy-pasted and edited) every time. After:
Use invoice-builder for Globex Corp, invoice 2026-0099, line items …, tax 7%.
That’s a parameter list. It’s also deterministic — the design system, the layout, the math, the styling are all baked into the skill. The model isn’t free-styling them anymore. That gap is the entire point of the platform. The user starts by chatting; the user ends by shipping skills.

​
Why it works

The skill-creator skill is itself an example of the same idea applied recursively: the work of “writing a skill correctly” used to be a long, careful prompt; now it’s a one-shot call to skill-creator. Every skill on the platform was built this way at some point — with a person staring at a prompt they had typed three times and asking “could this just be a function?” That’s the whole arc. Crawl (one-off prompts) → walk (the model uses live tools and sub-agents) → run (the user packages their own skills). Next: What’s next → — your skills become callable building blocks for n8n, CI, and any tool that can hit an HTTP endpoint.